What is VoxCPM2?
VoxCPM2 is a text-to-speech model from OpenBMB, the open-source group behind the MiniCPM language models, released in April 2026. It has 2 billion parameters on a MiniCPM-4 backbone, works without a discrete audio tokenizer, and outputs 48 kHz audio. OpenBMB says it was trained on more than 2 million hours of speech in 30 languages, plus nine Chinese dialects such as Cantonese and Sichuanese.
It does three jobs: it designs a new voice from a written description in brackets, clones a voice from a short clip (optionally steered with style instructions), and, given the clip's transcript as well, continues in that voice for the closest match. Weights and code are under Apache 2.0. The Hugging Face page records about 977,000 downloads in the past month, and the GitHub repository has about 41,000 stars.
Is VoxCPM2 worth it?
VoxCPM2 packs more into one commercially usable checkpoint than most open speech models: one 2B checkpoint covers 30 languages, voice design, cloning and streaming, and fits in about 8 GB of VRAM. Qwen3-TTS splits the same jobs across separate checkpoints, and most rivals with comparable language coverage carry non-commercial licences.
On benchmarks it is competitive rather than first. In OpenBMB's own Seed-TTS-eval table it scores a 1.84% English word error rate and 0.97% Chinese character error rate, ahead of VoxCPM1.5 (2.12% and 1.18%) but behind Qwen3-TTS on English (1.23%) and behind Meituan's LongCat-AudioDiT on every column. On CV3-eval, Fish Audio's S2 beats it in all but one language. Where it does lead is English voice design: on InstructTTSEval it scores 84.2, 83.2 and 71.4 against 82.9, 82.4 and 68.4 for Qwen3-TTS's VoiceDesign model. No independent arena such as Artificial Analysis lists it yet.
Speed is workable on one consumer card. The model card gives a real-time factor of about 0.30 on an RTX 4090, meaning ten seconds of audio take about three, and about 0.13 with the community Nano-vLLM engine. No time-to-first-audio figure is published. The ecosystem is wide: vLLM-Omni serves an OpenAI-style speech endpoint, and llama.cpp-omni, ONNX, Apple Neural Engine and ComfyUI ports exist.
The catch: cloning has no safeguards beyond a sentence in the model card forbidding impersonation and fraud. There is no watermark, no consent check and no detector. The MiniMax multilingual test also shows how uneven the 30 languages are: German scores a 0.68% word error rate, while Hindi scores 19.7%, Romanian 21.6% and Czech 24.1%. OpenBMB itself advises generating voice-design and style-control outputs one to three times to get a usable take.
Our take: for a team that wants to own its speech stack in European and East Asian languages, VoxCPM2 is the sensible default, and the money saved on licences should go on a native listener for each language you ship.
Where it earns its place
A games studio with a small audio team
Give 40 minor characters distinct voices without booking actors
Write a bracketed description per character, such as a gravelly older man speaking slowly, and render lines in batches on one RTX 4090, keeping the best of one to three takes. Apache 2.0 means no royalty on the shipped game.
A localisation agency dubbing training videos into German, Spanish and Japanese
Keep the original presenter's voice across languages
Pass the presenter's clip and its transcript for the closest clone, then generate each language. German and Spanish score well on OpenBMB's tests; budget a native reviewer, since error rates vary widely by language.
A backend engineer replacing a paid speech API in a reading app
Serve speech from an endpoint the app already knows how to call
Run vLLM-Omni, which exposes an OpenAI-compatible /v1/audio/speech route, on an 8 GB-plus GPU. The cost becomes fixed GPU rental instead of a per-character bill.
Before you commit
- Setup
- pip install voxcpm with Python 3.10 or later, PyTorch 2.5 or later and CUDA 12 or later; weights download from Hugging Face on first load. About 8 GB of VRAM in bfloat16.
- Integrations
- Python package, vLLM-Omni (OpenAI-compatible speech endpoint), Nano-vLLM-VoxCPM, llama.cpp-omni GGUF, ONNX, Apple Neural Engine and ComfyUI nodes. LoRA fine-tuning needs 5 to 10 minutes of audio.
- Data and privacy
- Self-hosted, so text and reference audio stay on your machines. Apache 2.0; the card forbids impersonation, fraud and disinformation and asks that AI audio be labelled, but nothing in the code enforces it.
- Cost at scale
- No licence fee and no first-party API. The cost is GPU time: at a real-time factor of 0.30, one RTX 4090 renders roughly three times faster than playback before batching.
- Limits
- No published time-to-first-audio. Maximum sequence of 8,192 tokens; the card warns of instability on very long or highly expressive inputs, and quality varies by language.
What’s good
- Apache 2.0 for weights and code, commercial use included
- Voice design, cloning and streaming in one 2B checkpoint
- About 8 GB of VRAM; real-time factor about 0.30 on an RTX 4090
- Leads Qwen3-TTS on English InstructTTSEval voice design
What’s not
- No watermark, consent check or misuse detector
- Behind Qwen3-TTS, LongCat-AudioDiT and Fish S2 on several error-rate tests
- Weak in some listed languages: 19.7% WER in Hindi, 24.1% in Czech
- Voice design and style control may need one to three tries per line
Best for
Product teams that need commercially usable open speech with voice design and cloning across European and East Asian languages on one GPU.
VoxCPM2 alternatives
VoxCPM2: common questions
- Is VoxCPM2 worth it?
- VoxCPM2 is the open speech model to start with when you need a commercial licence, many languages and voice design in one download. We rate it 4.2 out of 5.
- How much does VoxCPM2 cost?
- Free under Apache 2.0, commercial use included. OpenBMB sells no hosted API; the cost is your own GPU time.
- What are the downsides of VoxCPM2?
- No watermark, consent check or misuse detector; behind Qwen3-TTS, LongCat-AudioDiT and Fish S2 on several error-rate tests; weak in some listed languages: 19.7% WER in Hindi, 24.1% in Czech; voice design and style control may need one to three tries per line.
- Who is VoxCPM2 best for?
- Product teams that need commercially usable open speech with voice design and cloning across European and East Asian languages on one GPU.
- How much does VoxCPM2 cost for a team or in production?
- No licence fee and no first-party API. The cost is GPU time: at a real-time factor of 0.30, one RTX 4090 renders roughly three times faster than playback before batching.
- Is VoxCPM2 safe for company data?
- Self-hosted, so text and reference audio stay on your machines. Apache 2.0; the card forbids impersonation, fraud and disinformation and asks that AI audio be labelled, but nothing in the code enforces it.
- What are the best VoxCPM2 alternatives?
- The closest alternatives we have reviewed are Qwen3-TTS, OmniVoice, Fun-CosyVoice3 and Higgs TTS 3.
Answers checked Nov 7, 2026.
Make VoxCPM2? Show this rating on your site
